MINISFORUM MS-S1 MAX 128GB Max AI Compute Edition

The MINISFORUM MS-S1 MAX is an AI workstation powered by the AMD Ryzen AI Max+ 395, a Strix Halo processor that integrates a 16-core, 32-thread Zen 5 CPU, RDNA 3.5 graphics, and a next-generation NPU on a single package. The CPU reaches a maximum clock of 5.1 GHz, while the AI engine delivers up to 50 TOPS from the NPU and up to 126 TOPS of total platform AI performance across CPU, GPU, and NPU combined. Graphics come from AMD Radeon 8060S with 40 GPU cores running at 2900 MHz. This combination is built specifically for running large language models locally rather than sending data to a cloud service. The MS-S1 MAX pairs that compute with up to 128GB of unified LPDDR5x memory, dual 10GbE networking, USB4 V2 at 80Gbps, and a full-length PCIe x16 slot, in a 2.8 kg aluminium chassis with a built-in 320W power supply carrying CCC, FCC, UL, CE, and UKCA certification. It is designed to work either as a desktop workstation or as a node in a rack-mounted AI cluster.The MS-S1 MAX uses a unified memory architecture built on a 256-bit LPDDR5x bus running at up to 8000 MT/s, with a maximum capacity of 128GB. Because the CPU and GPU share a single memory pool rather than copying data between system RAM and dedicated VRAM, the system avoids the fixed VRAM ceiling that limits conventional discrete graphics cards, which is what allows models far larger than a typical consumer GPU can hold to run locally. The trade-off is that memory is soldered rather than socketed, so capacity is fixed at purchase and buyers should size their configuration for the largest models they intend to run. For storage, the system provides two M.2 2280 NVMe slots: one at PCIe 4.0 x4 and one at PCIe 4.0 x1, each supporting drives up to 8TB, for 16TB of total capacity with RAID 0 and RAID 1 supported. A slide-out chassis with an integrated board tray gives direct access to the storage and internal components for upgrades and maintenance without full disassembly.The MS-S1 MAX is designed to scale beyond a single unit, which is unusual for a desktop-class system. Two units can be clustered together, and in MINISFORUM testing a dual-unit cluster ran a 235B parameter model in Q4 quantization locally at an output speed of 10.87 tokens per second. The system also supports 2U rack deployment, where multiple units are cascaded into a distributed cluster for research teams or small studios. In MINISFORUM testing, a four-unit cluster ran the DeepSeek-R1 671B Q4 model, a 380GB workload in Q4_0 quantization, which is a model size normally associated with server-grade GPU deployments rather than compact workstations. A reserved cascade power-on header allows multiple units to be powered on and off together as a single managed group. Four switchable operating modes, Performance at 130W, Balanced at 95W, Quiet at 60W, and Rack, let the same hardware move between a quiet desk deployment and a sustained-load rack environment without hardware changes.Front connectivity on the MS-S1 MAX includes two USB4 ports at 40Gbps with DisplayPort 2.0 Alt Mode and 15W Power Delivery output, one USB 3.2 Gen2 port at 10Gbps, a 3.5mm audio combo jack, and dual digital microphones with AI noise cancelling for voice commands and calls. The rear carries two USB4 V2 ports at 80Gbps, two USB 3.2 Gen2 ports, two USB 2.0 ports, one HDMI 2.1 FRL output supporting 8K at 60Hz or 4K at 120Hz, and two 10GbE RJ45 ports using Realtek RTL8127 controllers. Because these are RJ45 rather than SFP+, they connect directly to a copper 10GbE switch with no transceiver required. Wireless is Wi-Fi 7 with Bluetooth 5.4. Cooling uses a pure copper heat spreader, a six heat-pipe array, phase change material in place of standard thermal paste, and dual turbine fans running at up to 3600 RPM, sustaining 130W under load with a 160W peak. The chassis measures 222.1 × 206.3 × 77.1 mm and weighs 2.8 kg, and supports both vertical desktop and horizontal rack orientation.Common use cases for the MINISFORUM MS-S1 MAX include running large language models locally through llama.cpp, Ollama, or LM Studio at parameter counts that exceed what a single consumer GPU can hold in VRAM, building and testing AI applications with low-latency on-device inference, multimodal model deployment, AI image and video generation, video editing and 3D rendering using the 128GB unified memory pool, and small-scale AI clusters for research teams and studios using 2U rack deployment. The full-length PCIe x16 slot running at PCIe 4.0 x4 adds industry-specific expansion such as capture cards, storage controllers, or specialized network adapters. Buyers typically choose the MS-S1 MAX over a discrete-GPU desktop when model size matters more than raw GPU throughput, since 128GB of unified memory holds models that would not fit in the VRAM of a comparably priced graphics card. Buyers who need upgradeable memory after purchase, or more than two NVMe slots, should consider the AI X1 Pro-370 instead, which uses replaceable SO-DIMM modules and three M.2 slots at lower memory bandwidth.